Research On Deep Image Denoising Network With Attention Mechanism

Image denoising algorithms aim to remove the noise generated during image generation or transmission,which can effectively improve image quality and are widely used in medical imaging,aerospace and security surveillance.The current deep learning-based image denoising algorithms have significantly improved the denoising performance,but the existing deep learning denoising networks have poor performance in image key information extraction and global feature information capture,making the texture details of denoised images still not clear enough.In addition,as the depth of the network increases,it is difficult to update the parameters of the shallow network,which makes the network training convergence difficult.To address the above-mentioned problems of existing deep learning denoising networks,this paper conducts research on deep learning denoising networks based on attention mechanism in terms of deep feature learning and network structure,which can further improve the performance of deep denoising networks and their practical application value.The main research of this paper is as follows:(1)A dense residual denoising network based on dynamic attention module is proposed.In response to the lack of autonomy in the attention mechanism selection in deep networks,this paper designs a dynamic attention module that uses channel attention and spatial attention to adaptively generate network weights,improving the network’s learning ability for key features.Meanwhile,an attention branch is constructed by adding the SE(Squeeze-and-Excitation)attention mechanism on the basis of the non-attention branch,forming a dual-branch structure.In addition,dilated convolution is introduced to increase the network’s receptive field,enabling more comprehensive acquisition of image features.Experimental results demonstrate that the proposed algorithm can effectively learn image features,improve the network’s representational performance,and preserve image details well.(2)A dense residual denoising network based on dynamic attention and LSTM(Long Short Term Memory)is proposed.To address the problem of poor correlation between feature information among image blocks and the difficulty of effectively utilizing shallow features in the network,this paper designs a dynamic attention LSTM module to perform sequence modeling and processing on images,thereby improving the correlation between image blocks.The overall network adopts a dense connection form to improve the utilization of network features.Experimental results demonstrate that the proposed algorithm can effectively restore image texture information even under high-intensity noise,achieving good visual effects.(3)A recurrent denoising network based on dual-layer residual LSTM is proposed.To address the problem of the network’s difficulty in effectively learning global image features,which leads to the network’s inability to effectively restore image texture information and training difficulties,this paper designs a dual-layer LSTM network.Firstly,the internal residual LSTM module is used to perform sequence modeling and learning on the feature image,and then this module is combined with a residual module to form a dual-layer LSTM structure to achieve global feature learning.Finally,using a recurrent connection form,the network’s various levels of feature information are fully utilized to improve the image denoising effect.Experimental results demonstrate that the proposed algorithm can effectively learn global image features and effectively remove noise under different noise levels.